Create submodels of an ExplainableBoostingRegressor(outer_bags=14)? about interpret

[paulbkoch]

Hi @gisandnes -- You can do either of these. If you want to train 14 separate ExplainableBoostingRegressor models with 1 outer bag each, you can combine them afterwards using the merge_ebms function. One complication with this route is that each of the separate EBMs will probably select a different set of interactions since there are slight variations in the interaction metrics across bags. You can resolve this by training EBMs without interactions, then using the measure_interactions function to choose a common set of interactions, then train new EBMs by specifying the interactions explicitly. There is an example in our documentation that shows how to handle custom interactions:
https://interpret.ml/docs/python/examples/custom-interactions.html

The other way, which I would recommend instead, is a bit hacker but easier. We store each of the bagged models inside the ExplainableBoostingRegressor object. You can force the EBM to behave as if it were one of the bagged models with:

# make the EBM behave as one of the bagged models
ebm.term_scores_ = ebm.bagged_scores_[3]
ebm.intercept_ = ebm.bagged_intercept_[3]
ebm.standard_deviations_ = None  # remove the error bars, which no longer apply with 1 bag

And it should behave identically to that individual bagged model for both prediction and feature/term importances.

I guess this functionality should be added to our API since it seems like it would be a common scenario.